Automated deep learning-based AMD stage detection in real-world OCT datasets

Abstract

Purpose : To develop an artificial intelligence (AI) method that can classify real-world optical coherence tomography (OCT) volumes and B-scans into different stages of age-related macular degeneration (AMD).

Methods : We trained a two-stage deep learning network to classify macula centered 3D volumes from Topcon OCT images into 4 different categories, i.e. Normal, Drusen, Atrophy and Neovascularization. A 2D ResNet50 was used to identify the disease categories on B-scans while an ensemble of models (ResNet18/Multilayer Perceptron) that uses the concatenated B-scan-wise output of the last hidden layer of the ResNet50 was used to classify the whole volume. Classification uncertainty estimates were generated with Monte-Carlo drop-out at inference time.
A total of 106,892 2D B-scans from a publicly available Spectralis OCT, Heidelberg Engineering, data set, supplemented with slides of atrophy, were used to pre-train a model in a transfer learning setup in order to ultimately classify Topcon OCT images from a data set consisting of 1,964 B-scans from 215 volumes of 189 eyes. The 3D ensemble model was trained on four binary classification problems with 5-fold cross-validation on a total of 1,575 OCT volumes from 1,418 eyes. Performance is reported as a balanced accuracy (ACC), precision (PREC), recall (REC), and specificity (SPEC).

Results : The Topcon scans were acquired from real world retrospective data sets extracted from the University Hospital Southampton and the Moorfields Eye Hospital as part of the PINNACLE study (75%) and supplemented with an internal data set (25%) from the Medical University of Vienna. The B-scan-level classification on a hold-out test set of 336 B-scans reached the performance of ACC/PREC/REC/SPEC: 0.97/0.97/0.96/0.99. The 3D classification of Topcon volumes achieved good performance on all classes on a hold-out test set out of 629 volumes, with high specificity in all classes (at least ACC/PREC/REC/SPEC 0.95/0.95/0.95/0.99) except for drusen (ACC/PREC/REC/SPEC 0.95/0.95/0.95/0.89).

Conclusions : The proposed approach enables the combination of different kinds of data sets into a domain specialized classifier that can reliably identify biomarkers associated with the relevant stages of AMD. Such automated AI tools can then serve to categorise patients by their disease stage, efficiently facilitating subsequent large-scale data analysis.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.